In sand casting production, process parameters significantly influence casting quality while complex 3D geometries create inherent quality control challenges. This paper presents a novel framework integrating process data analytics and 3D structural characterization for defect prediction in complex sand castings.
1. Process Data Acquisition and Analysis
Critical process parameters were systematically collected from foundry production systems:
| Process Stage | Key Parameters | Measurement Frequency |
|---|---|---|
| Sand Preparation | Moisture content, compactability, permeability | Hourly |
| Melting | Chemical composition, temperature | Per heat |
| Pouring | Pouring temperature, time, inoculation amount | Per casting |
The relationship between pouring temperature $T_p$ and defect occurrence follows:
$$P_d = \frac{1}{1 + e^{-k(T_p – T_c)}}$$
where $P_d$ represents defect probability, $T_c$ denotes critical temperature threshold, and $k$ is process sensitivity coefficient.
2. 3D Geometry Feature Extraction
A 3D Deep Convolutional Autoencoder (3D-DCAE) was developed for voxelized casting geometry analysis:
| Layer Type | Kernel Size | Output Shape |
|---|---|---|
| Input | – | 128×128×128 |
| 3D Conv | 5×5×5 | 64×64×64×32 |
| Max Pooling | 2×2×2 | 32×32×32×32 |
| Bottleneck | – | 512 latent features |
The reconstruction loss function combines binary cross-entropy and structural similarity:
$$L = \alpha \cdot BCE + (1-\alpha) \cdot (1 – SSIM)$$
where $\alpha = 0.7$ provides optimal balance between pixel-wise accuracy and structural preservation.
3. Hybrid Defect Prediction Model
The multi-modal prediction architecture combines process parameters $\mathbf{P}$ and geometric features $\mathbf{G}$:
$$y = \sigma(W_p \cdot \mathbf{P} + W_g \cdot \mathbf{G} + b)$$
where $\sigma$ denotes sigmoid activation, $W_p$ and $W_g$ represent parameter weights for process and geometry inputs respectively.
| Model | Training Accuracy | Test Accuracy |
|---|---|---|
| MLP | 92.6% | 86.1% |
| 2D-CNN | 93.9% | 90.7% |
| Proposed Model | 96.5% | 93.7% |
4. Industrial Validation
The framework demonstrated superior performance in sand casting production of complex components:
- Defect prediction F1-score: 0.92
- False positive rate: <5%
- Average prediction time per casting: 2.3s
Critical process-geometry interactions were identified through sensitivity analysis:
$$\frac{\partial y}{\partial P_i} = W_p^{(i)} \cdot y(1-y)$$
$$\frac{\partial y}{\partial G_j} = W_g^{(j)} \cdot y(1-y)$$
revealing significant correlations between section modulus variations and shrinkage defect formation.
5. Conclusion
This research establishes an effective methodology for quality prediction in sand casting through:
- Comprehensive process data collection and single-piece tracing
- Advanced 3D geometry characterization using deep autoencoders
- Multi-source data fusion with hybrid neural networks
The developed system enables proactive quality control in sand casting production, particularly for complex geometries prone to cold shuts, porosity, and shrinkage defects.